Ming-3.0-tts-0.5B vs Simba-3.2: When Free Weights Cost More Than $6.58

The pitch for open weights is that they are free. The arithmetic rarely survives the first GPU invoice. Ming-3.0-tts-0.5B is Ant Group's MIT-licensed speech checkpoint, downloadable at 2.96 GB and runnable on your own hardware. Simba-3.2 is SpeechifyAI's hosted voice, listed at $6.58 per million characters — the cheapest model in the top ten of the Artificial Analysis speech arena. At the arena's measured generation speed of 42.1 characters per second, producing a million characters with Ming takes roughly 6.6 hours on one GPU. If that GPU costs more than about a dollar an hour, the "free" model is the more expensive of the two. The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

That is the whole matchup in one line, but it is not the whole story, because the two models are not competing for the same job. The rest of this piece is about where the arithmetic holds, where it breaks, and the one board on which their scores are statistically indistinguishable.

Two very different products

Ming-3.0-tts-0.5B arrived without an announcement. Ant Group's inclusionAI published the weights to Hugging Face in September 2026 and the inference code to GitHub on October 19, 2026, under the MIT licence. It is a dense model on a Qwen2.5-0.5B backbone with a diffusion-transformer head, a bundled audio tokenizer and speaker encoder, 44.1 kHz output, seven languages, Sichuanese and Cantonese, 14 inline paralinguistic tags, voice cloning from a short clip, and voice design from a text description. There is no hosted endpoint and no price, because there is nothing to price.

Simba-3.2 is SpeechifyAI's hosted synthesis model, released July 1, 2026, with no open weights and no local option. It is the successor to Simba 3.0, which shipped February 19, 2026 at the identical $6.58 rate — a 117-Elo improvement for no increase in price, which is worth noting because it tells you how this vendor competes. SpeechifyAI does not publish a language count for Simba-3.2 that we could verify, so treat multilingual coverage as unconfirmed rather than assume it matches either the model above or below it in this comparison.

The practical difference in shape: one is an artefact you operate, the other is a service you call. Every comparison below is downstream of that.

The boards disagree, and the disagreement is the story

Artificial Analysis scores speech models two ways, and the two produce different answers here.

• Provider Voice, each model using its own native voices — Simba-3.2 Elo 1,247 (95% interval 1,234–1,260) at rank 9 of 98; Ming-3.0-tts-0.5B Elo 1,188 (1,173–1,203) at rank 19. A 59-point gap with no overlap.

• Controlled Voice, every model cloning the same eight reference voices — Simba-3.2 Elo 1,087 (1,074–1,100) at rank 10; Ming-3.0-tts-0.5B Elo 1,069 (1,053–1,085) at rank 14. An 18-point gap whose intervals overlap between 1,074 and 1,085.

That second line is the most useful measurement anyone has published about either model. Hold the voice constant and the difference between them is not statistically resolved. Ming's cloned voice is, on this evidence, roughly as good as Simba's. What separates them on the provider board is the voice cast each vendor ships: SpeechifyAI's native voices score well, and Ant's — 100-plus Chinese and five English — do not, which is consistent with the model card's own admission that speaker similarity trails content accuracy.

Ming's controlled-voice score also makes it first among open-weights models on that board, ahead of Fish Audio Drama 3 Preview at 1,059 and Breeze T​TS 2 at 1,019. So the accurate description is not "Simba sounds better." It is "Simba's default voices sound better; the engines are close, and the cloning path is close enough that the intervals touch."

The break-even, done honestly

The arena measures Ming at 42.09 characters per second of generation time and Simba-3.2 at 77.80, both on the same independent harness, one request at a time. That makes the comparison unusually tractable.

• A million characters on Simba-3.2 — $6.58, billed by SpeechifyAI, no infrastructure.

• A million characters on Ming at 42.09 characters per second — 23,758 seconds of generation, or about 6.6 hours.

• Break-even GPU price — $6.58 divided by 6.6 hours is roughly $1.00 per GPU-hour of generation.

Now the caveats, because this number is easy to abuse. Ming's release has been tested on an H800-80GB and an H20-96G and needs PyTorch 2.6 with CUDA 12.4; those are not dollar-an-hour GPUs, and the arena's harness hardware is not published. The 6.6 hours excludes model load time, failed generations, queueing, and the engineering hours to pin the environment and get text normalization working. Add those and the effective rate rises. If you already own idle capacity, the marginal cost is close to zero and the arithmetic flips entirely in Ming's favour; if you are renting by the hour for this workload alone, Simba is cheaper at almost any current cloud GPU rate.

The honest framing for a finance conversation is that Simba's $6.58 is a ceiling you can budget against and Ming's cost is a floor that grows with everything that goes wrong.

What the money does not buy

Price is where the hosted model wins. Capability is where the open one pulls back, and the differences are specific.

• Voice design — Ming generates a voice from a natural-language description with no reference audio. Simba-3.2 has no equivalent we could document. If your product needs a voice that does not exist yet, this alone decides it.

• Dialects — Ming handles Sichuanese and Cantonese given a reference clip in the dialect. SpeechifyAI publishes no dialect list for Simba-3.2.

• Paralinguistics — Ming ships 14 inline tags for laughter, sighs, coughs and hesitation in Chinese or English. Simba's tag vocabulary is not documented in a form we could verify.

• Portability — MIT weights mean no rate change, no deprecation notice, no vendor outage. They also mean no support, no SLA and no one to call.

• Operational risk — Simba is hosted, which is a feature until it is an incident; Ming is not hosted, which is a feature until the person who set up the environment leaves.

The model card is worth reading for the limitations alone, because they are unusually candid: multi-event paralinguistic sentences are unstable, emotion control degrades when the words carry no emotional signal, French similarity is the weakest language, and open-ended voice design does not match the strongest system it was evaluated against. Those are exactly the failure modes a demo does not show and a production pipeline finds in week three.

Where the routing layer fits, and where it does not

OrcaRouter does not host Ming-3.0-tts-0.5B or Simba-3.2, and this piece does not claim otherwise — Ming runs on your hardware, Simba through SpeechifyAI's API. The place a single key genuinely helps is around the voice rather than in it: the text model that writes the script, expands the abbreviations, or tags the emotion is the part of a speech pipeline that changes most often, and we route more than 200 of those at provider list price with no markup added. When a provider cuts a price, that cut is live on our side the same day, because we pass the list price through rather than repricing it. Automatic failover then keeps the script-generation half of the pipeline up while you decide whether the voice half should be a download or an invoice.

That is a narrower claim than "one API for everything," and it is the only one that survives a reader checking the catalogue.

The verdict, split by workload

Choose Simba-3.2 if the voice is a component of a product rather than the product itself, if you want a native voice that sounds finished without auditioning a cast, if your volume is modest enough that $6.58 per million characters disappears into the unit economics, or if you have nobody who wants to own a CUDA environment. The controlled-voice overlap is your reassurance that you are not paying a large quality premium for the convenience.

Choose Ming-3.0-tts-0.5B if you need a voice that does not exist in anyone's catalogue, if your audience is Chinese-first and dialect coverage matters, if your volume is high enough and your GPUs idle enough that the per-character meter is the expensive option, or if licence terms and data residency are non-negotiable. Go in knowing that you are adopting an unannounced research release with no support commitment and a documented list of weak spots, not a product.

And if you are genuinely unsure, the cheap experiment is to run both on a hundred sentences of your own text with the same eight reference voices and listen blind. That is roughly the protocol the controlled-voice board uses, it costs an afternoon, and on the published evidence the result is close enough that your content — not the leaderboard — will decide it.